Lecture 1 + Lecture 1 (IDK)

Chatbots and Goal-based dialogue agents

Talking for fun, therapy, mimicking informal humans, and keeping users engaged

ELIZA, Parry, Microsoft Tay, ChatGPT, XiaoIce

To help users achieve a task like booking flights, finding a restaurant, or getting a haircut

Google Assistant or Siri

A travel dialogue system that helps users book flights

A structure consisting of slots to be filled with values of a given type

ORIGIN CITY (type: city), question: “From what city are you leaving?”

ASR → NLU → Dialogue Manager → NLG → TTS

Controls the state and flow of the dialogue, handling user responses and system transitions

Determines who controls the conversation (system, user, or mixed)

System asks questions and controls the conversation

User directs the system, but the system can’t ask clarifications

Control of conversation shifts between system and user

Domain classification, intent determination, slot filling

Rule-based and Machine Learning

By filling templates with slot values (e.g., “X is a cheap restaurant in Chinatown”)

Ensuring mutual understanding between system and user

System directly asks for confirmation of user input (e.g., “Do you want to leave from Berlin?”)

System embeds confirmation in the next question (e.g., “When do you want to travel to Berlin?”)

The interactive functions of utterances (e.g., requesting, confirming, rejecting)

By detecting keywords, reflecting patient’s statements, and giving vague responses when no keyword is found

Rule-based are good for narrow domains but brittle; neural networks generate more flexible responses but may lack coherence

Base model pre-training + Reinforcement Learning from Human Feedback (RLHF)

Task success, efficiency (turns/time), human evaluation (enjoyment, naturalness)

A test where humans guess if they are speaking with a human or a machine

Purchase rates drop by 70% when users know they are talking to a bot

Rogerian philosophy (reflecting user’s statements back)

Actions performed by utterances: constatives, directives, commissives, acknowledgements

Paired utterances like question–answer or greeting–response

The process of drawing meaning beyond explicit words (e.g., implicature)

Inference guided by conversational maxims (Grice)

Study users/tasks, build prototypes (Wizard-of-Oz), iterative testing

Safety, verbal attacks, harmful stereotypes, conspiracy theories, privacy concerns

A dialogue where the system directs the conversation, and the user always knows what they can say next.

The system knows the expected words and topic, leading to better performance from ASR and NLU.

It is too limited and only suitable for very simple tasks.

A dialogue where the user directs the system, for example, by asking a question which the system then answers.

It cannot ask questions back or engage in clarification or confirmation dialogue.

The system asks questions to fill predefined 'slots' in a frame and performs a database query or action once the frame is complete.

It allows the user to answer multiple questions at once, avoiding strict constraints on the order of information provided.

They are good for narrow, scriptable applications.

They are brittle, meaning they can easily fail with unexpected user input.

The output can be a flexible combination of any words/characters from the training data.

They can struggle with staying coherent, being truthful, and often produce boring responses.

Task success, efficiency (e.g., number of turns), and quality (e.g., number of corrections).

Through human evaluation, asking participants about their experience (e.g., "How much did you enjoy talking to this chatbot?").

Automatic Speech Recognition.

Natural Language Understanding.

The process of identifying and extracting key pieces of information from a user's utterance to satisfy a task's requirements.